DynVision: A Toolbox for Biologically Plausible Recurrent Convolutional Networks

Abstract

AO_SCPLOWBSTRACTC_SCPLOWConvolutional Neural Networks (CNNs) trained for image recognition have demonstrated remarkable conceptual similarities to the primate ventral visual pathway, but their standard feedforward architectures lack the recurrent connections that are ubiquitous in visual cortex. Such recurrence is thought to underlie spatiotemporal phenomena including adaptation, delayed normalization, and robustness to noisy input. However, incorporating functionally beneficial recurrence into CNNs that captures spatiotemporal phenomena of biological vision remains challenging. Although recent advances have incorporated neurobiological constraints, the field lacks accessible tools for systematically comparing how different architectural choices, such as recurrence type, temporal delays, and connectivity patterns, shape neural dynamics and behavior. Here, we introduce DynVision, a modular open-source toolbox for constructing and evaluating biologically plausible recurrent convolutional neural networks (RCNNs). DynVision implements numerical ODE solvers with heterogeneous delays, supports five types of lateral recurrence ranging from simple self-connections to cortically-organized local recurrence, and separates scientific modeling decisions from implementation details through a configuration-driven design. Training is computationally efficient, achieving a 52% speedup over reference implementations. We demonstrate the framework through systematic exploration of the parameter space, revealing that qualitative differences in temporal dynamics are highly sensitive to often-implicit modeling choices such as the target location of recurrent integration and the temporal window used for loss computation. Critically, we find that continuous-time recurrent dynamics can naturally give rise to cortical temporal phenomena without requiring explicit divisive normalization, while a different recurrent configuration produces noise robustness approaching human-level performance. These findings suggest functionally distinct configurations of recurrence and highlight the challenge of creating fully realistic models, thus emphasizing the need for a comprehensive and cohesive modeling framework to aid exploration. Code and documentation are available at https://github.com/Lindsay-Lab/DynVision/.